Prompting guidance

Building effective fields requires making strategic decisions about field types and writing well-crafted prompts that extract accurate data from documents.

Choosing field types

Select field types based on document characteristics, data structure, and performance requirements.

Extraction vs. reasoning

Extraction field types provide more predictable results and more accurate source highlighting, especially for smaller documents.

Document reasoning fields might require more cleaning to format output consistently, and results are more variable. But, document reasoning fields perform better for large documents over 200 pages, even for straightforward fields like names or addresses. Document reasoning works particularly well when:

• Relevant information appears in multiple locations throughout the document.

• Contextual understanding or inference is required.

• Documents are dense or have complex formatting.

Writing effective prompts

Follow these general best practices to create clear, specific prompts.

• Keep prompts concise while including necessary detail.

• Use direct commands that state exactly what you want extracted.

  Extract the policy holder’s full name as it appears on the insurance declaration page.

• Provide context to distinguish between similar data points or explain document structure.

  Extract the effective date, not the application date, from the policy summary section.

• Specify formatting requirements only when basic standardization is required. For complex formatting, extract first, then clean.

  Extract the phone number in the format (XXX) XXX-XXXX.

• For list extraction fields, limit attributes to 10 for greatest accuracy (30 maximum).

• For document reasoning fields:

  • Use the Enhance prompt option to optimize clarity and effectiveness automatically.

  • Specify date and time if needed, otherwise the current date and time is used.

  • For chain of thought reasoning use numbered steps when you want to see the model’s logic. For a clean final result, use declarative commands instead.

• For derived fields:

  • Order fields so that referenced fields precede derived fields.

  • When referencing table or list extraction fields, normalize results as text using cleaning.

Extracting tables

Extract tables using either extraction or reasoning fields.

Follow these prompting best practices to extract tables from documents.

• Enable table recognition in digitization settings.

  To see recognized tables in a document, select the prediction icon in the header and enable Show detected objects for tables. Tables in the document are highlighted and you can use the adjacent table icon to view, copy, or download a table.

• Identify tables by descriptive characteristics rather than position or appearance.

  Extract the “Claims Summary” table that contains columns for Date, Description, Amount, and Status.

• Specify sheet name for multi-sheet Excel files.

  Extract the transaction table from the “Account Activity” or “Transaction History” sheet.

• Apply filtering and manipulation through prompt instructions.

  • Extract transactions and return results with amounts greater than $1,000.

  • Extract transactions and return results for 01 April through 15 April.

  • Extract transactions and sort amounts from smallest to largest.

• Specify output format when structured data is needed. Markdown is returned by default.

  • Extract transactions as JSON.

Extracting checkboxes, signatures, and barcodes

Extract information from certain visual objects, including checkboxes, signatures, and barcodes. The type of information you can extract differs by object type.

Extract objects using either extraction or reasoning fields.

For most object extraction, descriptive field names often provide adequate results:

• Filing Status (for grouped checkboxes)

• Signatory name, Signatory title, Signature date

• Barcode value

For standalone objects or more specific requirements, use prompts or descriptions:

• Is the filer claiming capital gains or losses? (standalone checkbox)

• Extract all signatures or Return yes if this document is signed

• Return all barcode values or Return yes if this document contains a barcode

Common patterns

These proven prompt patterns address frequent extraction scenarios typical of reasoning fields. Adapt them to your specific field requirements.

Chain-of-thought reasoning

Use numbered steps when you want to see the model’s reasoning process in the result.

1. Locate all medication entries in the prescription history section.
2. Extract the dosage for each medication.
3. Identify any medications exceeding 500mg dosage.
4. Return a list of medications exceeding 500mg with their dosages.

For a clean final result without intermediate reasoning, use declarative commands instead.

Identify all medications in the prescription history that exceed 500mg dosage and return only the medication names and dosages.

Date-based selection with fallback logic

Extract values associated with the most recent date, with fallback options when dates aren’t available.

Identify all lab results with collection dates.
Select the result with the most recent date.
If no dates are available, use the result that appears last in the document.
Return both the value and the collection date.

Conditional extraction

Extract data only when specific criteria are met, with clear handling of edge cases.

Extract the policy effective date only if:
- The date appears in the policy summary or declarations page.
- The date is within the last 2 years.
- The date format is valid (MM/DD/YYYY).
Otherwise, return "Date not found or invalid".

Multi-source data consolidation

Search multiple document sections in priority order and return the first valid result.

Determine the patient's current treatment plan:
1. Check "Treatment Plan" section for active interventions.
2. If not found, analyze "Provider Notes" for treatment recommendations.
3. Cross-reference with "Prescription History" to identify ongoing medications.
4. Synthesize findings into a summary of current treatment approach.
5. Return "No active treatment documented" if no information found.

Categorization

Categorize values using a predefined set of categories with clear criteria.

Determine the claim status as one of: [Approved/Denied/Pending/Under Review].
Base the determination on:
- Explicit status statements in the document.
- Decision language (approved, rejected, requires additional information).
- Presence of approval codes or denial reasons.
If status can't be determined, return "Status unclear".

Extraction with categorization

Extract numeric values and automatically categorize them into meaningful ranges.

Extract the patient's blood pressure reading and categorize:
- If systolic under 120 and diastolic under 80, return "Normal: [value]".
- If systolic 120--139 or diastolic 80--89, return "Elevated: [value]".
- If systolic 140+ or diastolic 90+, return "High: [value]".
- If blood pressure not found, return "Reading not available".

Aggregation and calculation

Perform calculations across multiple values found throughout the document.

Calculate total claim amount:
1. Locate all line items in the itemized charges section.
2. Extract the dollar amount for each line item.
3. Sum all amounts.
4. Return total in format: $X,XXX.XX.
5. If no line items found, return "No charges listed".

Presence detection

Determine whether specific information exists anywhere in the document.

Determine if pre-authorization is required:
1. Search for terms: "pre-authorization", "prior authorization", "pre-approval".
2. If found, return "Yes - pre-authorization required" with the specific requirement.
3. If explicitly stated as not required, return "No - pre-authorization not required".
4. If unclear, return "Pre-authorization requirement not specified".